Hi,
On Fri, Jul 1, 2022 at 8:07 AM Alexander Aring <aahringo@xxxxxxxxxx> wrote:
>
> Hi,
>
> On Thu, Jun 30, 2022 at 12:34 PM Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > On Thu, Jun 30, 2022 at 6:59 AM Alexander Aring <aahringo@xxxxxxxxxx> wrote:
> > >
> > > I send this patch series as RFC because it was necessary to do a kref
> > > change after adding __cond_lock() to refcount_dec_and_lock()
> > > functionality.
> >
> > Can you try something like this instead?
> >
> > This is two separate patches - one for sparse, and one for the kernel.
> >
> > This is only *very* lightly tested (ie I tested it on a single kernel
> > file that used refcount_dec_and_lock())
> >
>
> yes that avoids the warnings for fs/dlm.c by calling unlock() when the
> kref_put_lock() returns true.
>
> However there exists other users of kref_put_lock() which drops a
> sparse warning now after those patches e.g. net/sunrpc/svcauth.c.
> I think I can explain why. It is that kref_put_lock() has a release
> callback and it's _optional_ that this release callback calls the
> unlock(). If the release callback calls unlock() then the user of
> kref_put_lock() signals this with a releases() annotation of the
> passed release callback.
>
> It seems that sparse is not detecting this annotation anymore when
> it's passed as callback and the function pointer parameter declaration
> of kref_put_lock() does not have such annotation. The annotation gets
> "dropped" then.
>
> If I change the parameter order and add a annotation to the release
> callback, like:
>
> __kref_put_lock(struct kref *kref, spinlock_t *lock,
> void (*release)(struct kref *kref) __releases(lock))
> #define kref_put_lock(kref, release, lock) __kref_put_lock(kref, lock, release)
>
> the problem is gone but forces every user to release the lock in the
> release callback which isn't required and also cuts the API because
> the lock which you want to call unlock() on can be not part of your
> container_of(kref) struct.
>
> Then I did a similar thing before which would solve it for every user
> because there is simply no function pointer passed as parameter and
> the annotation gets never "dropped":
>
> #define kref_put_lock(kref, release, lock) \
> (refcount_dec_and_lock(&(kref)->refcount, lock) ? ({ release(kref); 1; }) : 0)
>
> Maybe a functionality of forwarding function annotation if passed as a
> function pointer (function pointer declared without annotations) as in
> e.g. kref_put_lock() can be added into sparse?
I think the explanation above is not quite right. I am questioning
myself now why it was working before... and I guess the answer is that
it was working for kref_put_lock() with the callback __releases()
handling. It has somehow now an additional acquire() because the
__cond_acquires() change.
Before the patch:
no warnings:
void foo_release(struct kref *kref)
__releases(&foo_lock)
{
...
unlock(foo_lock);
}
...
kref_put_lock(&foo->kref, foo_release, &foo_lock);
shows context imbalance warnings:
void foo_release(struct kref *kref) { }
if (kref_put_lock(&foo->kref, foo_release, &foo_lock))
unlock(foo_lock);
After the patch it's vice versa of showing warnings or not about
context imbalances.
- Alex
